500x Your Scientific Discovery With Sakana AI Scientist

Two years ago, I was part of a small research team working on a project called Wheat Disease Detection Using YOLOv8 and GAN.

It was exciting, challenging, and incredibly time-consuming. We spent almost an year building datasets, writing models, debugging, and testing. Every step took time, especially when we hit unexpected obstacles.

I remember late nights trying to figure out why our GAN was overfitting or why the YOLO model couldn't detect in certain lighting conditions in the field images.

We were proud of the final result, but the journey was long and slow.

Yesterday a video popped up in my feed about Google’s Co-Scientist, I got excited and questions started to unfold in my mind, "Will it help us boost our research, I bet it hallucinates, Is is resource and cost efficient?".

Everything we did in that year-long project- idea generation, literature review, experiment design, model testing, could have been done in under a day with the help of an AI scientist.

Don't believe me? See for yourself. In a recent talk, a Google researcher described a Friday afternoon where she had an idea for a study but needed to head home to take care of her kids.

Instead of dropping the idea, she plugged it into their AI co-scientist. By Monday morning, the system had created multiple hypotheses, refined them, tested different directions, and handed her a bunch of ready-to-explore approaches.

Today, companies like Google and Sakana aren’t just building AI agents that complete specific tasks for researchers, they’re creating AI systems that think, propose, test, and learn, much like a real scientist would.

In the upcoming years, AI will no longer be just a calculator or search engine it'll be a collaborator.

So why AI scientists?

Because they can speed up discoveries, reduce the cost of experiments, and even solve problems we once thought were impossible.

Whether it’s drug discovery, climate prediction, or quantum chemistry, AI co-researchers are helping us explore what was previously out of reach.

From my own experience, I know how slow and fragile research can be when we do everything manually.

That’s why I’m so fascinated and hopeful about what these AI scientists can bring to the table.

They're not replacing human creativity and insight, but they supercharge our potential to understand, discover, and build.

In this blog I will make you understand, Sakana’s AI Scientist, and why should you give it a try if already not.

Let me walk you through what they are, how they work, and what this all means for researchers like you and me.

What is Sakana AI Scientist?

After watching that video about Google’s Co-Scientist, I started digging deeper. That’s when I came across something even more mind-blowing, Sakana AI Scientist.

Their mission is simple: “To accelerate scientific discovery by automating the entire research lifecycle.” As someone who has spent nearly a year on a single research project, this caught my full attention.

I couldn’t stop thinking, what if the entire cycle, from idea to publication, could be done without me writing a single line of code or debugging a single script?

That’s exactly what Sakana is building.

Their AI Scientist doesn’t just assist. It actively generates hypotheses, tests them, learns from results, and improves over time.

I’m not talking about a chatbot that gives you paper summaries, I’m talking about an AI that can design experiments, analyze outcomes, and even write papers.

Their latest model AI-Scientist v2, wrote an entire research paper by itself, and it got accepted in a peer-reviewed journal. You can check it out online.

And they didn’t stop there.

Sakana recently partnered with researchers in a live experiment for ICLR 2025 on quantum computing.

They created a setup where scientists and AI can work together, just like co-authors, exploring new ideas, running simulations, and publishing findings. The code is all on GitHub, open for anyone to try.

How Sakana's AI Scientist Works?

When I first looked into Sakana’s AI Scientist, I didn’t expect much. I thought it might be just another chatbot that summarizes papers. But the more I explored, the more I realized—it’s not just answering questions. It’s doing full research. From idea to experiment to publication.

It All Starts with Ideas

Sakana’s AI begins by generating research ideas using something called Population-Based Training (PBT). Each idea is like a small experiment. It scores them based on how novel, interesting, or feasible they are.

Then it takes the best ones and improves them using Evolution Strategies (ES)—basically mixing and tweaking them like natural selection. Think of Darwin, but with code. It keeps only the strongest ideas and evolves them over time.

To make sure it doesn't waste time, it also uses Bayesian Optimization to fine-tune its parameters and focus on the most promising paths.

Running Experiments Without Help

Once the AI finds a good idea, it doesn’t wait for someone to pick it up. It uses a coding assistant called Aider to actually write the experiment code, fix bugs, rerun tests, and even plot the results.

It does this up to five times, refining the code and approach each time based on what it learns.

Yes, It Writes the Paper Too

When the results are ready, the AI writes the full research paper in LaTeX—Introduction, Methods, Results, everything. It adds charts, experiment logs, even real references from Semantic Scholar.

Then it revises and edits its own writing, making it cleaner, less repetitive, and easier to understand.

And Then... It Reviews Itself

It uses a second agent, trained on review guidelines from conferences like NeurIPS, to score itself on originality, clarity, and contribution. One of its papers, DualScale Diffusion, even scored high enough to meet the bar for top-tier AI conferences.

Soon, It Might Run Physical Labs Too

And this isn't where it stops. Sakana is now working on integrating with robotic labs. That means their AI Scientist might soon be able to run real-world experiments, not just simulations. We’re talking test tubes, sensors, chemistry gear all controlled by AI.

Conclusion

AI scientists like Sakana’s are helping us move faster from data to discovery. They can generate ideas, run experiments, and even write papers—but they still need human guidance.

As researchers, we need to stay involved. These tools are powerful, but it’s up to us to shape how they’re used. If we guide them well, they won’t replace us—they’ll help us do better science, faster.

FAQs

What are AI Co-Scientists?

AI co-scientists assist researchers by analyzing data, running simulations, and generating insights to speed up scientific discoveries.

How can AI help in scientific research?

AI can process vast datasets, identify patterns, and automate experiments, allowing scientists to focus on innovation and breakthroughs.

Can AI replace human scientists?

AI supports scientists but cannot replace them. It handles repetitive tasks and data analysis, while humans provide creativity and critical thinking.